Automatic Target Recognition in SAR Images Using CNN and Lee Filters

In this paper, an additional feature based convolutional neural networks (CNN) for synthetic aperture radar automatic target classification (SAR ATR) by using Lee filter will be presented. During the last decades, a lot of classical CNNs were proposed in order to target classification of SAR datasets, but the major problem of CNNs is that need in a large number of sample to train accurately. Also, images that are extracted from SAR usually contain a lot of noise. For these two problem, Lee filter will be used to obtain synthetic dataset of SAR data in order to increase the number of image to train the network better than classical CNNs and reduce the noise of SAR data to increase the accuracy of classification. Also, the proposed CNN includes three different steps. At first, more features by two kinds of CNNs by applying max-pool and average-pool subsampling operation will be extracted. Secondly, all of the information that are extracted from the two formed CNNs will be stacked into a single column vector in order to use both features in target classification. At the end, the proposed CNN by using stacked information and fully-connected layers will be trained. Also, the MSTAR dataset will be used to show the simulation result of the proposed method. By using the proposed method, 10 different classes can be recognized of military targets with overall classification accuracy of 98.88%.

Automatic Target Recognition in SAR Images Using CNN and Lee Filters Keywords:

convolutional neural networks , automatic target classification , Lee filter